MCMC for compositional spaces using CUDA GPUs
Cuda Python Shell Other
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Failed to load latest commit information.
data-sources
src
CHANGES.txt
LICENSE
Makefile
README.md
Shared.py
gpumcmc
make_primitives_header.py
postprocess.py
run-all.sh

README.md

GPUropolis

GPUropolis is software for using CUDA to run MCMC on compositional spaces. The current version implements a lightweight stack-based virtual machine in CUDA and runs a version of Metropolis-hastings on both the form of arithmetic expresssions (equations) and their constants as a CUDA kernel.

GPUropolis is under heavy development the computation and language lab at the University of Rochester.

Distributed under GPL3.

Input/output:

A standard run produces several files:

  • samples.txt - every chain's current state at the end of a block
  • MAPs.txt - every chain's best found hypothesis at the end of a block
  • log.txt - logging information, including the rng seed and parameters
  • performance.txt - performance statistics -- samples/second, function evals/second, etc.